BestJobFinder: an AI-native SaaS built with agent teams

BestJobFinder supports the whole job search, end to end. Users upload their CV, see jobs aggregated from several job boards, get an automatic match score for every job against their profile, apply and track their applications, prepare for interviews and keep track of offers.

The product has four frontends: a web application, an Electron desktop application, a Chrome extension and an admin panel, all built with React. They run on a Node.js microservice backend with separate API and worker services, designed for scalability and performance.

TypeScript is the main language on both sides, so code is reused across every layer. PostgreSQL with PgBouncer stores the data; Docker, Traefik and Let's Encrypt run the services; Prometheus and Grafana monitor them.

The features use prompt and context engineering, structured outputs and tool use, with model routing across Anthropic, Google, OpenRouter and local models served by Ollama and LM Studio (Llama 3.x, Qwen, DeepSeek, GLM and Kimi).

• Match scoring of every job against the user's profile.
• Tailored cover letters.
• CV improvement suggestions.
• Automatic answers to application form questions, drawn from the CV.

The whole development lifecycle runs on AI coding agents: Claude Code, Codex, GitHub Copilot and OpenCode. Each change moves through the same stages.

1. A written spec and an implementation plan come first.
2. Teams of specialized sub-agents implement the tasks in isolated git worktrees.
3. Separate agents review the work and test it in a real browser through Playwright over MCP.
4. A code knowledge graph gives the agents an accurate picture of the codebase.
5. Human review gates sit between the stages.

As founder and architect, I own the product, the architecture and the design and management of the agent teams that build it.